As companies add generative and agentic AI to customer operations, the most visible applications are often chatbots and virtual assistants. TDCX argues that some of the more consequential changes are happening behind the conversation, from quality assurance and knowledge retrieval to scheduling, summarization and the way frontline roles are designed.
That shift is part of the operating model behind TDCX’s new Foshan campus in China’s Greater Bay Area, which opened in September as the company’s sixth center in China. The facility launched with around 100 positions and is intended to support Cantonese, Mandarin and regional customer-experience operations with AI-enabled workflows.
In this TNGlobal Q&A, Michael Cowell, Managing Director of TDCX Hong Kong, discusses where AI is already changing CX delivery, why automation can increase the complexity of human work, which metrics are more useful than handling time, and what distinguishes an AI-native operation from one that has simply added AI tools.
TDCX describes the Foshan campus as AI-enabled. Beyond the technology at the site itself, where is AI already changing how customer-experience work is delivered across the region?
The most important changes are the ones customers never see. Chatbots and virtual assistants get a lot of attention, yet the bigger shift has happened behind the conversation.
Quality assurance is the clearest example. A traditional operation may review around 2 to 5 percent of interactions. With AI-assisted review, we can now look at effectively every contact. That changes what a supervisor does with their week. Instead of hunting for problems, they can coach against patterns the system has already surfaced. At Foshan this is built in from day one, with every pod using our tools across quality assurance, agent-assist workflows and voice-of-customer analytics.
Knowledge retrieval is another area. An agent handling a complex billing question used to hold the answer in their head or search through a portal while the customer waited. Now the relevant policy can surface in seconds. AI is also reducing administrative work around each contact, including summarization, disposition coding and follow-up drafting.
Which parts of CX work are being augmented most effectively today, and which still depend heavily on human judgment?
AI performs well when tasks are bounded and the source of truth is clear. That includes retrieval, summarization, translation drafts, quality review, forecasting and scheduling, as well as high-volume transactions such as order status or password resets.
What still needs a person is anything involving money, identity, emotion or ambiguity. A customer disputing a charge is not simply looking for information. They want a fair decision and someone who understands the situation. Knowing when to step outside the standard process, and when to escalate, remains human work.
There is also a second-order effect. As automation absorbs the simpler volume, the mix of contacts that reaches a person gets harder and cognitive load rises. Operations that automate easy work but keep the same staffing and training model can see quality slip. Success requires a broader transformation plan, not just a tool.
As generative and agentic AI become more capable, how do you expect frontline roles to change over the next two to three years? What new skills matter?
The volume of simple, scripted work will go down, while the complexity of what remains will go up. A new kind of role is also emerging around supervising the AI itself, including reviewing what it drafts, curating the knowledge it draws on and identifying where it is wrong.
Written reasoning, judgment and comfort working alongside the tools become more important. One of the most valuable frontline capabilities is being able to look at an AI-generated answer and quickly decide whether it is correct, needs adjustment or has to go to a human.
Productivity per person also rises. Teams tend to become leaner and more skilled rather than simply larger, and the work each person does becomes more valuable than the routine volume AI absorbs.
Multilingual delivery is central to regional CX. Where is AI improving language coverage, and where do local context, accents and cultural nuance still resist automation?
Written channels have improved significantly. Translating and maintaining a knowledge base across languages used to be a project. It is now closer to a workflow, and consistency can be better than when every market maintained its own version.
Voice is harder, and Cantonese shows why. Written standard Chinese and spoken Cantonese diverge significantly, and Hong Kong customers often code-switch into English in the same conversation. Speech recognition handles clean, single-language audio well but still struggles more with accented, code-switched, real-world speech.
Then there is register, which is not purely a language problem. How directly a request can be declined differs across markets such as Japan, Hong Kong and Australia. Getting that wrong can sound rude even when every word is technically correct. That is one reason we built Foshan around native Cantonese and Mandarin talent rather than relying on a translation layer.
What metrics should companies use to judge whether AI is actually improving outcomes? Is handling time still enough?
Handling time can become misleading. If automation removes short, simple contacts, average handle time can rise even while the operation becomes more efficient overall because the remaining cases are more complex.
We have built what we call a fair-score mechanism. Instead of judging an agent against one blended queue, analytics teams set baseline handling time and satisfaction expectations for different contact types, then weight performance by the actual mix a person handles.
I would give more weight to repeat-contact rate within seven days, cost per resolved issue rather than cost per contact, escalation rates in both directions, and satisfaction segmented by whether AI was involved. Containment also deserves skepticism. It tells you the customer left the automated channel, not necessarily that the problem was resolved.
How are enterprises approaching data security, privacy and residency when AI-enabled CX spans several markets?
The most common mistake is sequencing. Teams select a model or platform and then try to fit data governance around that choice. It should run the other way. First decide where each category of data is allowed to sit and which borders it can cross, then decide what technology touches it.
In practice, that means minimizing data by default, redacting personally identifiable information before inference when the model does not need it, setting clear retention limits on prompts and outputs, and maintaining audit trails that can reconstruct what happened.
Requirements across markets are not converging, so for multi-market operations we treat residency as a design constraint set with the client and its legal team rather than something standardized globally for convenience.
What makes Foshan and the Greater Bay Area useful as a delivery base as companies combine human talent with AI-enabled workflows?
The first factor is talent depth. Hong Kong has strong CX talent but not always enough of it at scale. The Greater Bay Area has a larger Cantonese, Mandarin and English-capable population that can support teams as clients grow.
Proximity matters as well. Foshan is about 90 minutes from Hong Kong Central, which makes it easier for client leaders to visit their teams. The campus opened with around 100 positions and room to grow, using a pod model where a Cantonese or Mandarin pilot can be live in roughly 60 to 90 days.
We are not suggesting clients move everything. Onshore Hong Kong talent still matters. Foshan adds another option for scale, speed and cost efficiency while keeping local-market knowledge close to the operation.
Looking ahead, what distinguishes a genuinely AI-native CX operation from one that has simply added AI tools to existing processes?
There is a simple test: if you switched the AI off tomorrow, what would break? In an operation that has added tools, things get slower. In one that is genuinely AI-native, the process stops because it was designed around the AI rather than added on top of it.
You can often see the difference in the organization before the technology. An AI-native operation has different staffing ratios, job descriptions and quality functions, and sometimes a different compensation model because volume handled is no longer the main measure of value.
The other marker is ownership. Where AI is treated as just a tool, technology owns the deployment while operations owns the result. Where it is native, the same team owns both. That is the model we are building around at Foshan. Most operations are still at the bolt-on stage, but the gap should become more visible in measures such as cost per resolved issue over the next few years.
Michael Cowell is Managing Director of TDCX Hong Kong. He has held leadership roles across North America, Europe and Asia in customer experience, outsourcing, hospitality and marketing communications, and previously served as Chairman of the Hong Kong Customer Contact Association.
Editor’s note: This Q&A has been lightly edited for clarity and TNGlobal house style. The substance of the interviewee’s responses has been preserved.
Share your perspective: TNGlobal welcomes contributed insights and expert commentary from across Asia’s technology and innovation ecosystem. Submit a contribution for editorial consideration, or explore more conversations in our TNGlobal INSIDER and TNGlobal Q&A and Interviews archive.
Mario Ferro on how GFE’s Pay It Forward model could scale education funding across ASEAN [Q&A]